Tech things that interest me. I founded QuantCDN.
Writing
- Your model may already ship a speculative decoder. Almost nothing loads it.Several popular open models carry a multi-token prediction head that most runtimes ignore. MTPLX loads them on Apple Silicon, and I have been widening the set of models it can take.
- The sortable IDs that stopped sorting on 14 August 2026opencode packs a 53-bit value into a 48-bit field, so its sortable IDs roll over every 795 days. The last rollover was in August. It showed up as three unrelated-looking UI bugs.
- One guard, 110× off the verify step. Speculative decoding on MLA models in MLX.Every MLA model in mlx-lm only used its fast attention path when there was exactly one query, which made speculative decoding slower than not speculating. Here is what the fix is and what it measured at.